Handle the case where a process is `SIGKILL`ed

[phargogh]

I have a case with the NCI work where my taskgraph-based pipeline is exhausting the available memory and linux's oom-killer is effectively kill -9ing the process. When this happens:

1. python starts up a new process to replace it, keeping the number of Processes the same as what the Pool expects at all times. The new process started has a different PID.
2. Because the Task was killed, the self.task_done_executing event is never set, which means that the graph hangs indefinitely.

The behaviorally consistent thing to do here might just be to terminate the graph when we detect that a process has been killed.

I'm labeling this as an enhancement because it's new behavior and because the workaround here is to just not exhaust the memory I'm allowed.